Primary vs Secondary Endpoint

A trial's primary endpoint is the one outcome its sample size was calculated on and its false-positive allowance was spent on. Everything else is secondary. Well-run protocols place key secondary endpoints in a fixed hierarchical testing order, so each is tested only if the one above it succeeded and the overall error rate is preserved; endpoints outside that hierarchy are exploratory whatever their p-value. Some trials define co-primary endpoints, in which case both must succeed.

In STEP 1 the primary endpoint was percentage change in body weight, with achievement of at least 5 percent loss as a co-primary, and the analysis of glycaemic and lipid measures sat below them. Cardiovascular outcome trials typically take a composite of cardiovascular death, myocardial infarction and stroke as primary and analyse the individual components as secondaries. The wording of an approved indication follows this structure closely: claims come from endpoints that were properly tested, not from every measurement collected.

This is what lets you judge a press release. A trial whose primary endpoint failed has produced no confirmatory finding, however striking a secondary looks, because the alpha was already committed and the trial was never sized for that comparison. The primary endpoint's confidence interval is the estimate to carry forward; secondaries are best treated as hypotheses for the next study.

The practical failure is outcome switching, where the endpoint reported in the paper is not the endpoint registered before enrolment, usually because something else looked better. The COMPare project documented this across trials in leading journals in 2016. A related move is presenting a secondary endpoint with a bare p-value and no mention that dozens of comparisons were run, which is multiplicity dressed as discovery.